Actually, AI Probably Isn't Taking Many Jobs, Yet

Headlines aside, there is arguably much less artificial intelligence job displacement than actually is the case, though some might argue for one particular nuance.

And that nuance is that AI is not actually displacing jobs when large enterprise job cuts are announced. Instead, a redirection of spending is envisioned, with the savings on labor being redeployed to support AI infrastructure creation. 

We can argue about whether that actually represents active AI substitution for existing labor, or is mostly a financial move designed to redirect resources to support a huge capital investment wave. 

source: Bret Jensen 

But the “fact” is that such job cuts are not really about AI displacing an existing job. For example, many firms overhired during the labor shortages caused by the Covid pandemic, and are now simply rebalancing. 

Some might call all of this "AI washing,” a strategic financial pivot disguised as an actual immediate substitution of AI for human labor:

• Capital reallocation: Companies cut headcount to free up the massive capital required for AI infrastructure, GPU compute, and model training

• Narrative: Frame a difficult, necessary financial correction as a visionary, tech-driven transformation.

• Efficiency: Reducing headcount in areas that were already overstaffed.

Company	Context of Announcement	Strategic Driver	Source
Meta	Workforce reductions linked to AI restructuring and operations simplification.	Reorganizing toward AI infrastructure; correcting over-hiring.	Financial Express
Amazon	Linked layoffs to efficiency measures and "AI-forward" operations.	Leaner management; funding heavy AI/cloud investment.	Medium
Block	Shrinking workforce citing AI tools accelerating productivity.	Reconfiguring to capitalize on strategic AI priorities.	CBS News
Pinterest	Cuts made to deliver on an "AI-forward" strategy.	Hiring new AI-proficient talent while reducing legacy roles.	CBS News
Cisco	"Hard decisions" made to shift investment toward AI era competitiveness.	Strategic discipline and focus on AI infrastructure.	NSJ Online
Klarna	Initially cited AI for headcount reductions; reportedly faced quality declines.	Experimentation with AI-led efficiency (with evidence of re-hiring).	Reddit (AI Tracking)

A reporting requirement by the state of New York whenever mass layoffs are conducted does not support the notion that AI is responsible for big layoffs there. 

source: Ahmed Abdelfattah

 

“In 2025, the New York Department of Labor updated the state’s Worker Adjustment and Retraining Notification Act system, asking businesses to disclose whether layoffs are related to artificial intelligence,” note Robert Quackenboss, Hinton partner and Michelle Meyer, Hinton associate. “In the year since the change took effect, no business has reported AI as a reason for layoffs.”

“Even though more than 160 different companies have filed WARN notices with the NYS DOL, not a single notice has attributed layoffs to AI technology or automation,” they say. 

To be sure, company resources are being diverted to capex and AI opex, and that is made possible by reductions of force. 

But AI is only an indirect cause, and not because it actually is being used to replace current human labor. 

The immediate cause is a need to invest resources in AI infrastructure. As with any firm resource allocation, there is a zero-sum element: what gets spent in one area means less spending somewhere else.

AI Impact: Many, Diffuse Winners; Some Concentrated Losers

It’s easy to empathize with recent college graduates concerned about the effect of artificial intelligence on entry-level jobs. At the moment, fewer such jobs (compared to past levels) seem to be a clear AI impact. 

On the other hand, it must also be noted that higher productivity is vitally important for any economy. 

Defined as the ability to produce the same output with fewer inputs; more with the same inputs or more output with fewer inputs, the point is that output, divided by inputs, does matter. 

And that matters because, over long periods, productivity growth is what raises living standards.

In other words, if one cares about higher living standards, one has to care about productivity. 

Productivity Raises Real Incomes

Without productivity growth, higher wages tend to create inflation rather than greater prosperity, as the same output then comes with higher input costs. 

Historically, the countries with the highest sustained productivity growth also achieved:

• higher wages

• shorter work weeks

• better healthcare

• more education

• longer life expectancy

• greater household consumption.

Consumers also benefit because productivity tends to reduce the real cost of goods and services:

• agriculture mechanization reduced food costs

• container shipping reduced transport costs

• semiconductor advances reduced computing costs

• software automation reduced transaction costs.

So luxuries often then become mass-market goods. In other words, productivity converts scarcity into abundance; expands consumer choice and product quality; saves time and reduces transaction friction. 

At the national level, productivity increases economic performance and all the benefits that flow from a robust economy. A society with fewer workers can still grow richer if each worker becomes more productive.

So, ultimately, productivity determines how much people can consume, save, invest, educate, heal, entertain, and invent. 

And the truth is that “productivity,” by definition, means wringing more output from any given level of inputs. And labor is among those inputs. 

So, again, it is easy to empathize with young college graduates who see AI as a threat. But productivity enhancers always represent such threats to someone, somewhere, by direct implication. 

The whole point is “producing more, using less.” 

But it is a somewhat nuanced issue. 

Productivity gains from AI and automation can hurt some suppliers of inputs, especially workers in tasks that are directly substituted, but they do not necessarily imply a net negative outcome for labor or other input suppliers on a societal level. 

Automation can displace some tasks and compress wages in some roles, to be sure.

However, if lower costs reduce prices enough, demand can rise, which can offset some or all of the labor reduction through expansion of output and new jobs. 

In other words, the key question is not whether a task is automated, but whether the market for the final product expands and whether labor is a substitute or a complement.

“Higher productivity” often, or virtually always, does mean benefits at the societal level. 

But some particular suppliers of inputs people) might not benefit in an immediate sense of protected, stable, well-remunerated employment, just as some suppliers of other physical inputs might suffer (buggy whip manufacturers after autos displaced horses for transport). 

Being a “Luddite” is not irrational. It is understandable. But resistance to powerful new automation technology of any type is ineffective in the long run. 

Productivity is overwhelmingly more powerful in the broader sense, in the long run, for the whole of society. 

In other words, there are many diffuse winners, but some concentrated losers.

AI Inference Pricing: Flat Rate, Usage, Hybrid

“Heavy users” are a recurring issue for any computing service using fixed rate billing.

For consumer or enterprise users, a general rule of thumb is that customers prefer predictable billing, and hence set rates. That might not be an issue if usage is predictable and moderate over time. 

But fixed rate billing can undermine a business model if usage and costs to provision are highly variable (marginal cost is high). 

On the other hand, fixed costs encourage sampling and experimentation, encouraging market growth. 

But there are customer hurdles when usage-based pricing is used. For suppliers, there is:

• Revenue volatility

• Usage spikes or dips can meaningfully swing revenue

• Sales friction. 

For buyers, the issues include”

• Bill shocks

• Complex buying process at scale

• Loss of cost prediction. 

Variable costs match “value” and “cost to supply,” as heavier users pay more. But that comes at the price of cost predictability, which tends to limit adoption. 

Fixed cost pricing (generally with usage allowances)  might include:

• Per-seat: $X per user/month

• Per-account: $Y per workspace or tenant/month

• Tiered plans: Basic, Pro, Enterprise (cost increases with tier)

Metered or usage-based billing might feature prices based on:

• Per API call: $0.001 per request

• Per token: $/1,000 tokens in or out

• Per compute hour: $/GPU-hour or vCPU-hour

• Per output unit: per generated document, per image, per transcription minute.

The perhaps-obvious compromise are hybrid models including both fixed cost and variable cost charging. For suppliers and customers alike, this provides some predictability of revenue or cost; plus some ability to match usage volume with cost. 

For most suppliers, usage-based billing makes most sense when:

• You supply infrastructure

• Usage and costs are highly variable

• Customers are developers and customers experimenting

• Cash and margin discipline are top of mind.

Fixed pricing will tend to work better when:

• AI is a feature in a broader SaaS product

• Usage per customer is relatively consistent or capped

• Buyers want simple prices for yearly contracts

• Model and cloud costs are low and covered at the plan price.

Hybrid appeals when:

• You want predictable baseline revenue and upside from power users

• You have meaningful unit costs but strong sales

• You sell into finance or procurement buyers and need to protect margin.

Hybrid is generally preferable, if only because AI inference costs are highly variable and scale directly with usage. 

Also, as a practical matter, hybrid helps assure some level of ability to plan for recurring revenue magnitudes, per account or per user. 

And most suppliers will sell to consumers, developers, smaller and enterprise-sized customers, some preferring low entry costs; others preferring cost predictability at scale. 

And a reasonable assumption is that usage will climb over time, for virtually every customer set. 

All that noted, many prior computing products have shifted from flat rate to metered, or metered to flat rate. The shift to cloud computing involved another change from “own” to “rent,” with the emergence of subscriptions (flat rate, generally) and end of perpetual license (one payment, upfront). 

And where “subscription” is the product payment structure, flat rate tends to be preferred, where possible. That tends to be the case where costs are not highly variable. 

The issue with AI inference is simply that costs are fairly linear with usage. 

Also, there should eventually be a stronger trend towards "billable units" that align with standard business workflows. In healthcare, that might be cost per-patient, per-test, or per-episode. 

In other businesses, perhaps the shift is to per-user and per-task metrics. There also is much talk of pricing based on outcomes or value, but that always requires clear metrics, always a difficult task. 

Product/Shift	Direction	Effects on Volume/Use Cases
AWS Cloud (2006+)	To usage-based (pay-as-you-go) from traditional on-prem/capex	Massive increase in volume; enabled elastic scaling, burst workloads, startups, and new experiments. Lowered barriers, exponential growth in compute/storage usage.
Adobe Creative Suite to Creative Cloud	Perpetual/one-time to subscription (flat recurring)	Increased customer engagement, continuous updates, higher lifetime value; broader access but some backlash from users preferring ownership. Steady revenue growth.
Microsoft Office/Enterprise Software	Perpetual to subscription (e.g., Office 365)	Higher retention, regular updates; shifted to predictable opex. Increased adoption in teams via per-user flat elements; some optimization of usage.
Mobile Data Plans	Usage-based/metered to flat-rate "unlimited" (then back with caps)	Flat-rate caused surge in data consumption, app innovation (streaming, social); re-introducing usage elements controlled heavy users without killing growth much.
Snowflake/Databricks (cloud data)	Usage-based (compute credits)	Enabled pay-for-query model; grew adoption for variable analytics workloads; high volume in data-heavy use cases with efficiency focus.
GitHub Copilot (2026)	Flat-rate/subscription to usage-based	Ongoing; expected to better align costs with heavy AI coding use; may gate extreme users but improve sustainability and model choice.

SpaceX IPO Estimates $26.5 Trillion AI Addressable Market

Not the sort of company mission one normally sees in an S-1 (initial public offering) filing!

On the other hand, the expected future revenue leans on artificial intelligence, not space or connectivity. 

So such goals as “colonizing Mars” are essentially the sizzle on the steak. 

SpaceX S1 filing 

SpaceX estimates a total addressable market of $28.5 trillion:

• $370B in "Space" (space-enabled solutions including Starship (fully reusable booster), lunar economy, point-to-point Earth transport, in-orbit manufacturing)

• $1.6T in Connectivity (Starlink Broadband ~$870B + Starlink Mobile ~$740B, plus enterprise/government)

• $26.5T in AI (compute infrastructure (terrestrial + orbital), subscriptions, advertising, enterprise applications, Grok models, consumer/enterprise/government adoption, X platform monetization. 

Year	Total Revenue	Starlink/Connectivity	Launch/Space	AI/xAI
2025 (Actual)	~$18.7B	~$11.4B	~$4.1B	~$3.2B
2026 (Est.)	~$20–28.5B	Strong growth (e.g., $18–20B+)	Stable/moderate growth	Growing but capex-heavy
Longer-term (2030+)	$100B+	Dominant	Scaled via Starship	Major contributor

Google, Blackstone Create TPU "as a Service" Business

Google and Blackstone’s TPU-as-a-service venture is important for any number of reasons:

• it turns TPUs from a mostly Google-hosted product into a broader external infrastructure platform

• strengthens Google’s push to monetize its custom silicon

• gives AI customers a non-Nvidia acceleration path

• might clarify the neocloud business model. 

Blackstone is committing $5 billion in equity and an initial 500 MW of capacity coming online in 2027. 

The move tends to ratify the GPU as a service market and provides an alternative to the Nvidia ecosystem, at least in the “bare metal” portion of the business. 

The venture also might intensify pricing pressure and reduce differentiation in the inference market. 

The venture also tests the durability of the neocloud business itself. Today, a global scarcity of high-end AI training and inference compute creates the basis for the market.

Neoclouds originally emerged as stopgaps to address the GPU shortage, but their bare-metal economics are fragile, being based on what most believe are temporary shortages of capacity. 

Perhaps their long-term viability hinges on their ability to move up the stack into AI-native services, which puts them in direct competition with hyperscalers. And some will note how little protection the business has, given the thin profit margins and high continuing capital investment. 

source: McKinsey 

Neoclouds have a strong demand story, but their business model is structurally difficult because they combine very high capital intensity with fast hardware depreciation and aggressive price competition. The result is a market that can grow fast while still being hard to make sustainably profitable.

The core problem is that graphics processing units are expensive, and their resale or rental value falls quickly as new generations arrive. 

McKinsey notes that over a typical five-year depreciation horizon, GPU-hour pricing can decline by half or more, which forces providers to recover capital quickly or risk stranded assets.

So neoclouds must keep raising capital to buy the next wave of chips even while the prior fleet is losing value. This makes cash flow, financing terms, and utilization rates far more important than simple revenue growth.

GPU clouds are not just chip businesses; they are power, cooling, networking, and operations businesses as well. High energy costs, high-density racks, and increasingly complex cooling requirements raise operating expense and add execution risk.

Up to this point, neoclouds are heavily dependent on Nvidia for the chips, networking ecosystem, and much of the software stack.

Google will test that thesis.

A big reason neoclouds emerged was that they could undercut hyperscalers on price and provisioning speed, sometimes by large margins. But hyperscalers are responding.

That means the initial “GPU scarcity arbitrage” is not a durable moat by itself.

The strategic tension is that investors often want neoclouds to move up the stack into managed services, orchestration, inference platforms, or sector-specific solutions. Those layers can improve retention and margins, but they also bring neoclouds into direct competition with hyperscalers that have deeper ecosystems and broader product bundles.

So the firms face a hard choice: stay close to bare-metal GPU rental, where margins are thin, or build higher-value services, where competition is tougher and sales cycles are longer.

That suggests a need to pioneer niche markets, such as sovereign compute and specialized workloads.